Medical Vision Systems has developed a new image segmentation technology that improves the accuracy and robustness of automated medical image segmentation. Our commercial goal is to offer a product that produces accurate, fast, fully automatic segmentations at low cost to aid the Intensity Modulated Radiation Therapy (IMRT) treatment of prostate cancer. The use of Intensity-Modulated Radiation Therapy (IMRT) has become the preferred method for treating prostate cancer through radiation doses in excess of 70 Gy [1]. This technology has been developed to ensure that therapeutic doses of radiation are delivered only to the target organs, the prostate and the seminal vesicles, without affecting nearby non- involved structures, such as the bladder and the rectum. This technique relies on dose planning software and dynamic multileaf collimators to shield these sensitive organs. Without preventative measures, 15% to 35% of patients developed grade 2 or worse rectal toxicity [2, 19, 20, 22, 31, 32] and, with less statistical certainty, increased late urinary complications [5, 12, 30]. IMRT Treatment Planning software requires accurate segmentations of the organs in the abdomen and pelvis. Segmentations of the prostate that include neighboring tissues can irradiate those tissues unnecessarily, while segmentations of organs such as the bladder or rectum that include too much surrounding tissue can interfere with complete delivery of radiation dose. Thus, an inaccurate segmentation has a real effect on the quality of life of the patient. Furthermore, the time required to produce a manual segmentation of the organs of interest significantly limits the number of radiation therapy treatments that can be undertaken. In addition, if image-guided radiation therapy (IGRT) supplants IMRT in the same way that IMRT has supplanted conformal therapy, then manual segmentation will become an even more limiting bottleneck in contemporary radiation therapy. Preliminary work has begun through partnerships with three cancer treatment centers. The investigators have built upon previous work that shows the Auto Context Model (ACM) with AdaBoost is effective at subcortical segmentation in magnetic resonance (MR) imaging [15, 14, 16, 17, 18]. We propose to build upon this foundation to develop a new learning-based image segmentation system capable of accurately and automatically segmenting organs of interest in abdominal x-ray computed tomography (CT) scans taken at different imaging facilities. Accurate segmentation of organs in abdominal CT images is complicated by large variations in abdominal anatomy, motion of abdominal tissues, the limited contrast between anatomic structures in x-ray computed tomography, and variations in imaging equipment, protocols, and techniques. In this Phase I SBIR proposal, we will address the limitations of existing segmentation tools to achieve the accuracy and automation required for IMRT treatment planning. We propose to develop an innovative learning-based segmentation system using both bottom-up and top-down approaches. We will construct a novel feature dictionary, implicitly incorporate atlas-based segmentation methods through the use of image registration techniques, and build a modest "ground truth" database using images acquired and segmented by our collaborators. A Phase II proposal would demonstrate the clinical feasibility of these techniques, addressing both the stability of the system across multiple imaging sites and determining the impact of the proposed system on patient outcomes through increased segmentation accuracy. PUBLIC HEALTH RELEVANCE: Medical Vision Systems has developed a new learning-based image segmentation technology that can simultaneously improve the accuracy and robustness of automated medical image segmentation. Our commercial goal is to offer a product that produces accurate, fast, and fully automatic segmentations at low cost to aid in Intensity Modulated Radiation Therapy (IMRT) and, in the future, Image Guided Radiation Therapy (IGRT) for prostate cancer.